¥0.32 for a 100,000-token translation batch

01 / The ¥0.32 receipt is a calculation, not a live price

The cheapest translation model is not necessarily the one with the lowest input-token price. Translation produces substantial output, and output charges can reverse the ranking. A useful comparison starts with a receipt: source tokens, translated tokens, each applicable rate, and the total yuan actually charged.

Here is a worked comparison for a batch containing 100,000 input tokens and producing 120,000 output tokens. Every rate below is hypothetical, not a current provider or CheapRelay quote. These numbers demonstrate the arithmetic; they are not measured performance results or evidence that one public model translates better than another.

02 / Two quotes, one identical translation batch

Call the competing quotes A and B. Both charge separately for input and output, with rates expressed in yuan per million tokens. Assume no cache activity, retries, extra billable tokens or minimum charges in this first pass.

• ■ Quote A input: 100,000 tokens × ¥1.00 per million = ¥0.10.
• ■ Quote A output: 120,000 tokens × ¥2.00 per million = ¥0.24.
• ■ Quote A total: ¥0.34.
• ■ Quote B input: 100,000 tokens × ¥2.00 per million = ¥0.20.
• ■ Quote B output: 120,000 tokens × ¥1.00 per million = ¥0.12.
• ■ Quote B total: ¥0.32.

On this receipt, B ranks first despite charging twice as much for input. Its batch costs ¥0.02 less, a 5.88% reduction relative to A. At 1,000 identical batches, the totals become ¥320 and ¥340: a ¥20 difference before any operational overhead.

03 / Rank accepted source text per yuan, not output volume

For this controlled example, B processes approximately 312,500 source tokens per yuan; A processes approximately 294,118. That is a useful price-only ranking because the source workload is fixed. Rewarding output tokens per yuan would encourage a bad outcome: longer translations could appear more economical simply because they contain more tokens.

Across real public models, even source-token counts need care. Tokenizers differ, so the same document can generate different counts. Keep the source documents identical, then rank accepted documents or source characters per yuan. Use each model's actual billed tokens to calculate spending, rather than assuming every model sees the same 100,000-token input.

04 / A 10% retry load turns the winner into the loser

Cheap first attempts do not guarantee cheap usable translations. Suppose B requires extra runs adding 10% to its original token spending, while A needs none. B's adjusted bill becomes ¥0.352, above A's ¥0.34. At 5% extra spending, B costs ¥0.336 and remains narrowly cheaper.

The break-even point is 6.25% additional spending on B, assuming A stays unchanged. Human correction can overwhelm the difference even faster: adding ¥0.05 of review cost to B's original batch raises its effective cost to ¥0.37. Track terminology errors, broken placeholders, omitted sentences and review effort separately from the API receipt.

05 / Put public model names beside verified receipts

DeepSeek V4 Flash and Gemini 2.5 Flash are two candidates to test, not declared winners here. No verified current rates or translation test results were supplied for either, so assigning them the illustrative A and B prices would manufacture a ranking.

Instead, run the same representative documents through both. Record the exact model identifier, test date, input and output usage, cache reads, cache writes, other billable usage, and actual yuan charged. Check accuracy and formatting against the same acceptance rules. Then sort by total spending per accepted source document, including retries and review.

06 / Keep the relay bill inside the comparison

CheapRelay provides pay-as-you-go access through one key, with no monthly fee or subscription. Billing follows real token usage, including cache reads and writes, so those receipt lines belong in your comparison rather than being treated as free.

Sign up for CheapRelay to run a small, capped translation comparison before committing a larger batch. The defensible winner is the model that delivers acceptable translations for the lowest verified total—not the one attached to the smallest advertised input rate.